Atmospheric waveguide prediction method and system based on non-uniform feature fusion and graph deduction

Abstract

The invention belongs to the field of marine atmosphere remote sensing, and provides an atmospheric waveguide prediction method and system based on non-uniform feature fusion and graph deduction, and the method comprises the steps: carrying out the dynamic topological feature coding according to a radar echo gradient change rate and meteorological feature disturbance intensity, and obtaining a multi-modal reconstruction feature; corresponding graph structures are constructed according to different scales of the multi-modal reconstruction features, and the edge weight of each scale graph structure is calculated to form an atmospheric waveguide graph; in the atmospheric waveguide diagram, determining whether to interact according to the edge weights of the nodes of the to-be-interacted scale and the nodes of other scales, and for all the nodes of the to-be-interacted scale and the nodes of other scales, which are determined to interact, performing cross-scale interaction based on a gating mechanism to generate an interactive atmospheric waveguide diagram; and carrying out modal alignment on the two modals in the interactive atmospheric waveguide diagram, and carrying out adaptive aggregation according to the confidence coefficient between the two modals to obtain an atmospheric waveguide prediction feature diagram. According to the invention, high-precision non-uniform atmospheric waveguide inversion prediction is realized.

Description

Technical Field

[0001] This invention belongs to the field of marine atmospheric remote sensing technology, specifically relating to an atmospheric waveguide prediction method and system based on non-uniform feature fusion and graph extrapolation. Background Technology

[0002] The statements in this section are merely background information related to the present invention and do not necessarily constitute prior art.

[0003] To achieve over-the-horizon communication and target detection at sea, atmospheric waveguide environment prediction is a crucial scientific problem. Under the coupling effect of the ocean and atmospheric environment, due to the nonlinear interaction between temperature and humidity gradients and atmospheric disturbances, atmospheric waveguides exhibit significant non-uniformity and abrupt changes in their spatiotemporal distribution, posing a severe challenge to accurate prediction. Currently, typical non-uniform atmospheric waveguide prediction methods have limitations in several aspects, failing to meet practical application requirements. These limitations mainly manifest in insufficient small-scale feature extraction capabilities, deficiencies in modeling complex morphologies and capturing long-range dependencies, and difficulty in effectively distinguishing regions with similar features. Specific limitations are as follows: Insufficient small-scale feature extraction: Existing models have difficulty effectively preserving and enhancing weak small-scale features caused by local non-uniformity in deep network structures, resulting in a significant decrease in the predictive ability of fine waveguide structures in complex ocean and atmospheric backgrounds.

[0004] Limited morphological modeling capabilities and insufficient long-range dependency capture: Existing methods struggle to flexibly adapt to the complex and varied geometries of non-uniform atmospheric waveguides, and their modeling capabilities for irregular waveguide structures are insufficient. Furthermore, the models lack effective mechanisms to capture long-range spatiotemporal correlations between waveguide features over large scales, limiting the global consistency of predictions.

[0005] The current algorithm has a deficiency in feature discrimination capability: it is not good at distinguishing features of atmospheric waveguide regions with highly similar and gradual changes in shape, which affects the completeness and accuracy of the prediction results and is prone to misjudgment or omission.

[0006] In existing related technologies, prediction algorithms based on SP-URNet employ 3D convolutional autoencoders and temporal convolutional networks (TCNs) for spatiotemporal feature modeling. However, they lack a dedicated dynamic enhancement mechanism for non-uniform small-scale features, and their long-range correlation capture relies on a fixed-structure network, making it difficult to adapt to the irregular dynamic changes in waveguide morphology. Another type of method borrows from graph convolutional networks in vehicle fine-grained classification, inferring the correlation between features through graph structures. However, the graph structures used are usually statically constructed, failing to reflect the dynamic and heterogeneous nature of atmospheric waveguide features, and also lacking refined modeling and contrastive learning mechanisms for edge features. Summary of the Invention

[0007] To address the aforementioned issues, this invention proposes an atmospheric waveguide prediction method and system based on non-uniform feature fusion and graph inference. This invention integrates non-uniform feature dynamic enhancement, fine-grained graph structure adaptive inference, and edge feature comparison learning, aiming to specifically address the shortcomings of existing methods in small-scale feature preservation, complex morphology modeling, and similar region feature differentiation, thereby achieving high-precision non-uniform atmospheric waveguide inversion prediction.

[0008] According to some embodiments, the first aspect of the present invention provides an atmospheric waveguide prediction method based on non-uniform feature fusion and graph derivation, employing the following technical solution: Atmospheric waveguide prediction methods based on non-uniform feature fusion and graph derivation include: Based on the radar echo gradient change rate and the intensity of meteorological disturbances, dynamic topological feature encoding is performed to obtain multimodal reconstruction features; Based on the different scales of the multimodal reconstruction features, corresponding graph structures are constructed, and the edge weights of the graph structures at each scale are calculated by integrating spatial distance and feature similarity to form an atmospheric waveguide graph. In the atmospheric waveguide map, whether to interact is determined based on the edge weights of the nodes at the scale to be interacted with nodes at other scales. For all nodes at the scale to be interacted with that are determined to interact with nodes at other scales, cross-scale interaction is performed based on a gating mechanism to generate an interactive atmospheric waveguide map. Mode alignment is performed on the cross-scale interaction features of the two modes in the interactive atmospheric waveguide map, and adaptive aggregation is performed based on the confidence between the two modes to obtain the atmospheric waveguide prediction feature map.

[0009] According to some embodiments, a second aspect of the present invention provides an atmospheric waveguide prediction system based on non-uniform feature fusion and graph derivation, employing the following technical solution: An atmospheric waveguide prediction system based on non-uniform feature fusion and graph derivation includes: The dynamic topology feature encoding module is used to perform dynamic topology feature encoding based on the radar echo gradient change rate and the intensity of meteorological feature disturbances to obtain multimodal reconstruction features; The layered graph construction module is used to construct corresponding graph structures based on different scales of multimodal reconstructed features, and to calculate the edge weights of graph structures at each scale by integrating spatial distance and feature similarity to form an atmospheric waveguide graph. The cross-scale graph inference module is used to determine whether to interact with nodes of the target scale and nodes of other scales in the atmospheric waveguide map. For all nodes of the target scale and nodes of other scales that are determined to interact, cross-scale interaction is performed based on a gating mechanism to generate an interactive atmospheric waveguide map. The feature calibration and aggregation module is used to perform mode alignment on the cross-scale interactive features of two modes in the interactive atmospheric waveguide map, and to perform adaptive aggregation based on the confidence between the two modes of cross-scale interactive features to obtain the atmospheric waveguide prediction feature map.

[0010] According to some embodiments, a third aspect of the present invention provides a computer-readable storage medium.

[0011] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps in the atmospheric waveguide prediction method based on non-uniform feature fusion and graph deduction as described in the first embodiment above.

[0012] According to some embodiments, a fourth aspect of the present invention provides a computer device.

[0013] A computer device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the steps in the atmospheric waveguide prediction method based on non-uniform feature fusion and graph derivation as described in the first embodiment above.

[0014] According to some embodiments, a fifth aspect of the present invention provides a computer program product or computer program.

[0015] A computer program product or computer program includes computer instructions stored in a computer-readable storage medium, wherein a processor of a computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computer device to perform the steps in the atmospheric waveguide prediction method based on non-uniform feature fusion and graph derivation as described in the first embodiment above.

[0016] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention effectively represents the detailed features of non-uniform atmospheric waveguides by studying a detailed representation algorithm based on non-uniform feature fusion, significantly improving the prediction accuracy of atmospheric waveguides. Utilizing fine-grained perceptual map inference, it can effectively learn long-range atmospheric waveguide features and handle the heterogeneity between atmospheric waveguide features of different granularities. The application of edge feature extraction and contrastive learning methods further enhances the prediction performance in local non-uniform regions, while also strengthening the network's robustness.

[0017] This invention can preserve small-scale features and solve the problem of deep network feature suppression in existing SP-URNet technology through a dynamic adaptive fusion module, thereby improving the representation ability of small-scale non-uniform waveguides in complex backgrounds. Moreover, it can also adapt to complex shapes, breaking through the limitations of static graph structure in deformable graph convolution, and accurately capturing the irregular shape and long-range correlation of atmospheric waveguides. It can also distinguish similar regions by solving the problem of distinguishing highly similar gradient regions through Gaussian modeling and two-stage contrastive learning, thereby improving the completeness and accuracy of prediction. Attached Figure Description

[0018] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0019] Figure 1 This is a flowchart of an atmospheric waveguide prediction method based on non-uniform feature fusion and graph derivation in an embodiment of the present invention; Figure 2 These are comparison curves of key indicators for atmospheric waveguide prediction using different methods in embodiments of the present invention. Detailed Implementation

[0020] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0021] It should be noted that the following detailed description is illustrative and intended to provide further explanation of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0022] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the scope of exemplary embodiments according to the invention. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0023] Where there is no conflict, the embodiments and features in the embodiments of the present invention can be combined with each other.

[0024] Example 1 like Figure 1As shown, this embodiment provides an atmospheric waveguide prediction method based on non-uniform feature fusion and graph derivation. This embodiment uses the application of this method to a server as an example for illustration. It is understood that this method can also be applied to terminals, and can also be applied to systems including terminals, servers, and other components, and can be implemented through interaction between the terminal and the server. The server can be an independent physical server, a server cluster composed of multiple physical servers, or a distributed system. It can also be a cloud server providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network servers, cloud communication, middleware services, domain name services, CDN security services, and big data and artificial intelligence platforms. The terminal can be a smartphone, tablet, laptop, desktop computer, smart speaker, smartwatch, etc., but is not limited to these. The terminal and server can be directly or indirectly connected via wired or wireless communication, which is not limited herein. In this embodiment, the method includes the following steps: Step S1: Based on the radar echo gradient change rate and meteorological feature disturbance intensity, perform dynamic topological feature encoding to obtain multimodal reconstruction features; Step S2: Construct corresponding graph structures based on different scales of multimodal reconstruction features, and calculate the edge weights of graph structures at each scale by fusing spatial distance and feature similarity to form an atmospheric waveguide graph; Step S3: In the atmospheric waveguide map, determine whether to interact based on the edge weights of the node at the scale to be interacted with nodes at other scales. For all nodes at the scale to be interacted with that are determined to interact, perform cross-scale interaction based on the gating mechanism to generate an interactive atmospheric waveguide map. Step S4: Perform mode alignment on the two modes of cross-scale interaction features in the interactive atmospheric waveguide map, and perform adaptive aggregation based on the confidence between the two modes of cross-scale interaction features to obtain the atmospheric waveguide prediction feature map.

[0025] In a specific embodiment, in step S1, dynamic topological feature encoding is performed based on the radar echo gradient change rate and the intensity of meteorological feature disturbances to obtain multimodal reconstruction features. The process is as follows: Step S1.1: Calculate the radar echo gradient change rate and meteorological feature disturbance intensity based on the radar echo map and meteorological observation dataset; Step S1.2: When the radar echo gradient change rate exceeds the electromagnetic characteristic threshold or the meteorological characteristic disturbance intensity exceeds the physical characteristic threshold, determine to initiate topology reconstruction; Step S1.3: Extract the electromagnetic feature tensor of the radar echo image and the physical feature tensor of the meteorological observation dataset, and generate a topological weight matrix after processing by a spatiotemporal attention mechanism; Step S1.4: Generate multimodal reconstruction features using a parameterized convolution kernel generator function based on the topological weight matrix and the adjusted node connection relationships.

[0026] First, it should be noted that the existing SP-URNet prediction algorithm uses a fixed convolutional structure to extract features, which easily loses small-scale non-uniform information. This embodiment designs a bidirectional complementary adjustment module and a multi-scale asymmetric complementary modulator to overcome the limitations of the fixed structure. It uses a dual-stream characterizer to extract local and global features separately, and uses dynamic adaptive weights to couple local salient features with global semantics to suppress background interference.

[0027] First, in constructing the dynamic topology feature encoding module for non-uniform atmospheric waveguides, an innovative topology reconstruction mechanism based on multi-mode threshold triggering was designed. This mechanism monitors the gradient change rate of radar echoes in real time. Disturbance intensity of meteorological data To construct a dual-sensitive dynamic response system.

[0028] In step S1.1, calculations are performed based on the radar echo map and meteorological observation dataset to obtain the radar echo gradient change rate and the intensity of meteorological characteristic disturbances, including: Acquire radar echo maps and meteorological observation datasets; The raw radar echo data contains information such as radar echo intensity and spatial distribution, and is stored in the form of radar echo maps (two-dimensional / three-dimensional matrix); Raw meteorological data includes physical parameters of the ocean and atmosphere such as temperature, humidity, air pressure, and wind speed, and is stored in the form of meteorological observation datasets (spatiotemporal series or spatial grid data).

[0029] The radar echo gradient change rate is obtained based on the radar echo map, and the intensity of meteorological feature disturbances is obtained based on the meteorological observation dataset. Specifically, these two parameters are derived using the following formula: Radar echo gradient change rate The spatial gradient calculation based on radar echo maps is given by the following formula:

[0030] in, For spatial coordinates, Echo intensity.

[0031] Meteorological disturbance intensity The formula for calculating time-series fluctuations based on raw meteorological data is as follows:

[0032] in, For real-time temperature, Average temperature; For real-time specific humidity, The average specific humidity.

[0033] In step S1.2, when the radar echo gradient change rate exceeds the electromagnetic characteristic threshold or the meteorological characteristic disturbance intensity exceeds the physical characteristic threshold, topology reconstruction is initiated, including: when Exceeding the preset electromagnetic characteristic threshold When this occurs, it indicates a drastic change in the electromagnetic wave propagation environment within the radar detection area, possibly corresponding to a sudden change in waveguide layer thickness or an abnormal refractive index gradient; while when Exceeding the physical property threshold When either of these conditions is triggered, it signifies a significant disturbance in the spatiotemporal distribution of atmospheric physical parameters (such as temperature, humidity, and air pressure), indicating a potential morphological evolution of the waveguide structure. These two thresholds correspond to the critical change points of electromagnetic and physical properties, respectively, and together constitute the quantitative index system for waveguide inhomogeneity. Secondly, when either of the above conditions is triggered, the topology reconstruction process is immediately initiated. This design allows the coding module to accurately capture the dynamic evolution characteristics of non-uniform atmospheric waveguides while maintaining computational efficiency, effectively solving the problem of insufficient adaptability to environmental changes in traditional static architectures. The gradient change rate of the radar echo is used as an example. Disturbance intensity of meteorological data As a trigger condition, when or ( , When the electromagnetic property threshold and physical property threshold are met, topology reconstruction is initiated.

[0034] In step S1.3, the electromagnetic feature tensor of the radar echo image and the physical feature tensor of the meteorological observation dataset are extracted to generate a topological weight matrix, including: Extract electromagnetic feature tensors from radar echo maps and physical feature tensors from meteorological datasets; Specifically, the electromagnetic feature tensor, which includes the radar echo gradient rate of change, is first extracted from the radar echo map. Spatial distribution information; extracting physical feature tensors from raw meteorological data, which include the intensity of meteorological feature disturbances. Spatial distribution information; Based on the radar echo gradient change rate in the electromagnetic feature tensor and the meteorological feature disturbance intensity in the physical feature tensor, a multilayer perceptron is used to reconstruct the features and obtain the topological weight matrix.

[0035] Secondly, the electromagnetic characteristic tensor in In the physical characteristic tensor As a key sensitive parameter, it is input into the multilayer perceptron (MLP) and generated as a topological weight matrix after softmax normalization. This matrix contains the geometric prior information of the atmospheric waveguide structure at the current moment, and each element represents the contribution of the corresponding spatial region to the predicted target. (Topological weight matrix) The structure is as follows:

[0036] in, for The topological weight matrix at time step [time]. For multilayer perceptrons, This is the normalization function.

[0037] It is important to note that each element in the topological weight matrix represents the contribution of a spatial region to the predicted target, and this contribution is calculated using both electromagnetic and physical feature tensors; the elements in the matrix are ordered based on feature distance.

[0038] In step S1.4, multimodal reconstructed features are generated based on the topological weight matrix using a parameterized convolutional kernel generation function, including: This embodiment proposes an adaptive convolution kernel generation mechanism based on spatiotemporal topological weights. This mechanism achieves accurate modeling of complex atmospheric environments by dynamically adjusting the local receptive field structure of the convolution operation. The parameterized convolution kernel generation function is as follows:

[0039] in, It serves as the basic convolution kernel, responsible for extracting general features of atmospheric waveguides; The dynamic offset is determined by the topological weight matrix. Element-wise modulation is applied to the kernel. This design enables the convolution kernel to adjust its receptive field shape and parameter distribution according to real-time changes in the atmospheric environment, automatically increasing the sampling density in regions of abrupt waveguide gradient changes, and maintaining efficient computation in uniform regions.

[0040] It should be noted that the multimodal reconstruction feature consists of one radar reconstruction feature and one meteorological reconstruction feature, which are then combined together.

[0041] In a further embodiment, in step S2, corresponding graph structures are constructed according to different scales of the multimodal reconstruction features, and the edge weights of the graph structures at each scale are calculated by fusing spatial distance and feature similarity to form an atmospheric waveguide graph. The process is as follows: Step S2.1: Divide the multimodal reconstruction features into microscale, mesoscale, and macroscale according to their different scales; Step S2.2: Fuse the spatial distance and feature similarity of multimodal reconstruction features at different scales, and calculate the edge weights of the graph structure at the corresponding scale; Step S2.3: Based on the edge weights of graph structures at different scales, construct graph structures at different scales, using each multimodal reconstruction feature as a node and the relationship between nodes as edges, and use the graph structures at different scales to form atmospheric waveguide maps.

[0042] To construct a hierarchical graph: Based on the node set optimized by dynamic topological feature encoding, the nodes are divided according to scale, the edge weights of each scale are calculated, and finally "micro / medium / macro scale subgraphs" are formed. In addition, a graph neural network is used to dynamically adjust the node connection relationship so that the model architecture can adapt to the current waveguide morphology. The nodes are multimodal reconstruction features at various scales (i.e., the feature vector corresponding to each spatial location in the multimodal reconstruction features output in step S1). The edges refer to the association between nodes. The weight of the edges is calculated by fusing spatial distance and feature similarity. The larger the weight value, the stronger the waveguide feature association between the two nodes.

[0043] First, the layered graph construction mechanism aims to overcome the limitations of traditional single-scale graph models in capturing multi-scale feature associations. By accurately dividing scale boundaries and dynamically defining edge weights, it achieves refined modeling of waveguide features at different spatial scales.

[0044] In step S2.1, the multimodal reconstruction features are divided into micro-scale, meso-scale, and macro-scale based on their different scales, including: Specifically, firstly, preliminary optimization is performed on the global topological relationships of multimodal reconstruction features (such as removing redundant feature units and strengthening node associations in highly sensitive regions). At this point, a scale-based graph structure has not yet been constructed, with the aim of providing a high-quality set of nodes for subsequent hierarchical graph construction.

[0045] Based on the physical characteristics and spatial distribution patterns of atmospheric waveguides, all multimodal reconstruction features are divided into three levels according to their influence range: Microscale ( (0-50 meters) mainly corresponds to the local fine structure of the evaporation waveguide, such as the region of abrupt change in refractive index gradient within 10-30 meters near the sea surface. These features have a small spatial scale but have a significant impact on the low-altitude propagation of electromagnetic waves. Mesoscale ( The area (50-500 meters) encompasses the distribution of surface waveguides, including non-uniform structures of medium range such as frontal transition zones and eddy current disturbances, and serves as a key link connecting microscale details with macroscale background. Macroscale ( A range of 500 meters or more represents a large-scale atmospheric disturbance background, such as a pressure system or airflow vortex at the hundred-kilometer level, whose dynamic evolution directly regulates the formation and dissipation of small- and medium-scale waveguides.

[0046] In step S2.2, the spatial distance and feature similarity of multimodal reconstructed features at different scales are fused, and the edge weights of the graph structure at the corresponding scale are calculated, including: The edge weights of the graph at each scale are calculated using the following formula:

[0047] in, For nodes and Spatial distance, For feature cosine similarity, The attenuation coefficient is... This is the summation symbol.

[0048] To accurately characterize the correlation strength of waveguide features at various scales, an edge weighting mechanism that integrates spatial distance and feature similarity was designed for each scale map: edge weights It is determined by two parts, one of which is the spatial decay term. ,in, For scale Next node and Spatial distance, The first is the scale-adaptive attenuation coefficient (larger values ​​at microscales to strengthen local correlations, and smaller values ​​at macroscales to preserve long-distance dependencies); the second is the feature consistency term. The correlation between the physical properties (such as refractive index and temperature gradient) of two nodes is quantified by calculating the cosine similarity of the eigenvectors. The product of the two is normalized to form the final edge weight, which ensures that spatially adjacent waveguide units are preferentially correlated, while avoiding correlation deviations caused by ignoring feature similarity due to simple distance measurement (such as waveguide structures in different locations but with the same physical properties).

[0049] This multi-scale graph construction method enables microscale graphs to focus on capturing small-scale non-uniform abrupt changes (such as thickness jumps in evaporating waveguides), mesoscale graphs to accurately characterize the continuous evolution of waveguide morphology within a region, and macroscale graphs to effectively correlate large-scale background perturbations, laying a structured foundation for subsequent cross-scale attention flow feature interactions, thereby solving the problem of blurred non-uniform feature representation caused by scale aliasing in traditional models.

[0050] In step S2.3, based on the edge weights of graph structures at different scales, each multimodal reconstruction feature is used as a node, and the relationships between nodes are used as edges to construct graph structures at different scales. Atmospheric waveguide maps are then formed using these graph structures, including: Each multimodal reconstruction feature at different scales is used as a node; The edges of graph structures at different scales are determined by the relationships between nodes at different scales; Construct map structures at various scales, including microscale maps, mesoscale maps, and macroscale maps; Atmospheric waveguide maps are obtained by stitching together the map structures at various scales in order of scale.

[0051] After calculating the edge weights of the graphs at each scale, the graph structures at each scale are constructed, with multimodal reconstruction features as nodes and the relationships between nodes as edges. For the first Scale node and The edge weights are used to form micro-scale, meso-scale, and macro-scale graphs; Edge weights can also guide cross-scale interactions. As prior information on the strength of node associations, the gating mechanism combines Determine whether cross-scale interaction is needed (e.g.) (Larger nodes are prioritized for interaction), please refer to step S3 for details.

[0052] In step S3, in the atmospheric waveguide map, whether to interact is determined based on the edge weights of nodes at the target interaction scale and nodes at other scales. For all nodes at the target interaction scale and nodes at other scales that are determined to interact, cross-scale interaction is performed based on a gating mechanism to generate an interactive atmospheric waveguide map, including: In the construction of layered maps of non-uniform atmospheric waveguides, cross-scale attention flow is the core mechanism for achieving organic coupling of features at micro, meso, and macro scales. It precisely controls the interaction intensity of features at different scales through gating mechanisms, solving the problem of multi-scale correlation breaks caused by scale isolation in traditional layered models. The non-uniformity of atmospheric waveguides is essentially the result of multi-scale dynamic coupling: macro-scale atmospheric circulation (such as cyclones at the 100 km scale) regulates the movement of mesoscale fronts (50-500 meters) through pressure gradients, while abrupt changes in temperature and humidity at mesoscale fronts directly affect the thickness and refractive index gradient of microscale evaporation waveguides (0-50 meters). Therefore, single-scale feature modeling alone cannot capture this cross-level causal relationship; cross-scale attention flow is necessary to achieve directional transfer and dynamic fusion of features.

[0053] Step S3.1: In the atmospheric waveguide diagram, if the edge weights of nodes at the scale to be interacted with nodes at other scales exceed the weight threshold, then the two nodes will interact. Specifically, the design of the gating mechanism follows the principle of prioritizing importance: For each pair of interactions, the scale (e.g., macro scale) and mesoscale Mesoscale With microscale Macroscale With microscale ), Gating mechanism based on edge weight Dynamic judgment is used; for example, if the correlation strength between macro-scale nodes and micro-scale nodes exceeds a weight threshold, then macro-scale is activated. With microscale Interaction (such as when macro-scale strong pressure disturbances directly affect micro-scale evaporation waveguides).

[0054] Step S3.2: For all nodes at the determined interaction scale and nodes at other scales, calculate the corresponding cross-scale interaction features based on the gating mechanism, and combine them to generate an interactive atmospheric waveguide map, including: Cross-scale interaction features, interacting through a gating mechanism, are as follows:

[0055] in, It is a gated function. For feature splicing, For the first Nodes at scale Multimodal reconstruction features, It is the first Nodes at scale Multimodal reconstruction features, , It refers to cross-scale interaction features. In other words, the cross-scale interaction features here include radar cross-scale interaction features and meteorological cross-scale interaction features.

[0056] After cross-scale interaction, the cross-scale interaction characteristics of different nodes in the map structure at different scales are obtained, and combined to generate an interactive atmospheric waveguide map.

[0057] Specifically, the gating function based on feature similarity and physical relevance is as follows:

[0058] in, , These are feature vectors at two different scales. This is a feature splicing operation.

[0059] The gating function extracts cross-scale feature association patterns through convolutional layers, and then... Activation generates weight coefficients in the 0-1 interval, which quantify the contribution of source-scale features to target-scale modeling. For example, when the macroscale... When strong pressure disturbances occur, the gating function will be equal to or equal to the mesoscale... The interaction of frontal features is assigned a high weight (close to 1) to ensure the efficient transmission of key perturbation information; while at microscale... When local noise exists (such as random fluctuations in radar echoes), the gating function will automatically reduce its mesoscale orientation. The transfer weight is close to 0 to avoid noise accumulation.

[0060] The interaction process of cross-scale attention flow exhibits a "two-way feedback" characteristic: On the one hand, macro-scale background constraints (such as atmospheric stability parameters) are channeled to the meso-scale through gating, providing global priors for frontal evolution; the regional features at the meso-scale are then channeled to the micro-scale through gating, guiding the local fine modeling of evaporation waveguides.

[0061] On the other hand, sudden changes at the microscale (such as a sudden increase in the thickness of the evaporation waveguide) will also be fed back to the mesoscale through gating, correcting the gradient parameters of the front; the anomalous features at the mesoscale will be further fed back to the macroscale, updating the evolution model of atmospheric circulation.

[0062] This two-way gating interaction ensures that macroscopic constraints guide microscopic modeling in a directional manner, and enables dynamic correction of the macroscopic model by microscopic mutations. This allows the layered diagram to adaptively capture the cascading effects of non-uniform atmospheric waveguides "from global to local" and the feedback adjustment "from local to global", ultimately improving the model's overall ability to represent complex waveguide structures.

[0063] In a specific embodiment, in step S4, mode alignment is performed on the two modes of cross-scale interaction features in the interactive atmospheric waveguide map, and adaptive aggregation is performed based on the confidence between the two modes of cross-scale interaction features to obtain the atmospheric waveguide prediction feature map. The process is as follows: Step S4.1: Map the two modes of cross-scale interactive features in the interactive atmospheric waveguide map to the same high-dimensional space. When the maximum mean difference between the two modes of cross-scale interactive features in the high-dimensional space meets the difference threshold, the aligned radar features and aligned meteorological features are obtained. Step S4.2: Calculate the information entropy of the cross-scale interaction features of the two modes in the interactive atmospheric waveguide map, determine the confidence level of the cross-scale interaction features of the two modes based on the information entropy of the cross-scale interaction features of the two modes, and adaptively fuse the aligned radar features and aligned meteorological features according to the confidence level of the cross-scale interaction features of the two modes to obtain the atmospheric waveguide prediction feature map.

[0064] First, mode alignment is the core step in resolving the discrepancy between the distribution of electromagnetic features and meteorological physical features in radar echoes. It quantifies and minimizes the distribution differences between the two modes in a high-dimensional feature space, providing a consistent feature base for subsequent fusion. The non-uniformity of atmospheric waveguides is often reflected in both electromagnetic propagation and the physical environment: electromagnetic features such as radar echo power attenuation and Doppler shift are essentially indirect responses to the atmospheric refractive index gradient; while meteorological data such as temperature, humidity, and air pressure are direct physical quantities that determine the refractive index distribution. However, there are significant differences in the modal characteristics of the two—electromagnetic features are easily affected by propagation path disturbances (such as multipath effects), exhibiting strong nonlinear fluctuations; physical features are closer to the intrinsic properties of the atmosphere, but are limited by observation density (such as sparse distribution of meteorological stations), and may have spatial sampling biases. This modal gap between indirect responses and direct physical quantities, if not calibrated, can lead to an imbalance in feature weight allocation during fusion, and even introduce spurious correlations (such as misjudging radar noise as waveguide abrupt changes). To achieve accurate alignment, the maximum mean difference (MMD) is used as a measure of distribution distance, mapping the two modal features to the same high-dimensional space, thus transforming the nonlinear distribution difference into a quantifiable linear distance.

[0065] Step S4.1: Map the two modes of cross-scale interactive features in the interactive atmospheric waveguide map to the same high-dimensional space. When the maximum mean difference between the two modes of cross-scale interactive features in the high-dimensional space meets the difference threshold, the aligned radar features and aligned meteorological features are obtained. The process is as follows: Specifically, the reconstructed radar electromagnetic features and reconstructed meteorological physical features in the interactive atmospheric waveguide map are mapped to the same high-dimensional space through a mapping function (Gaussian kernel function, or other mapping functions can be selected, which are not specifically limited here). Modal alignment calculates the maximum mean difference (MMD) between two modes in a high-dimensional space. Specifically, it is used for radar cross-scale interactive feature sets in interactive atmospheric waveguide maps. Interactive feature set with meteorological scale Through mapping function (e.g., Gaussian kernel function) transform them into high-dimensional features respectively. and Then calculate the maximum mean difference between the two, and adjust the parameters of the mapping function until the maximum mean difference between the two meets the difference threshold, so as to obtain the aligned radar features and aligned meteorological features. The parameters of the mapping function are adjusted based on minimizing the difference loss function, which is calculated by subtracting the mean vector of the two values. Norm square, i.e.

[0066] By minimizing the difference loss function This ensures that the statistical distributions of the two modes are consistent; the loss function directly quantifies the degree of deviation between the two modes in the global distribution. When the electromagnetic anomalies of radar echoes spatially match the physical disturbances in meteorological data, A smaller value indicates good modal alignment; When there is a deviation (such as radar capturing waveguide signals but meteorological data not reflecting the corresponding temperature and humidity changes). As the value increases, the driving model adjusts the feature mapping parameters through backpropagation until the distribution trends of the two modalities become consistent.

[0067] This mode alignment mechanism is specifically designed for complex scenarios involving non-uniform atmospheric waveguides: in abrupt regions such as passing fronts, it strengthens the correlation between electromagnetic features and temperature / humidity gradient distributions to avoid misjudgments caused by single-mode noise; in uniform regions such as open sea surfaces, it weakens redundant alignment constraints, preserving the unique information of each mode. Ultimately, mode alignment provides a co-distributed feature basis for subsequent adaptive fusion, enabling the detailed fluctuations of radar echoes and the macroscopic trends of meteorological data to work synergistically in a unified feature space, significantly improving the robustness of non-uniform waveguide predictions.

[0068] Secondly, based on feature entropy ( The modal confidence calculation mechanism is the core of achieving dynamic fusion of heterogeneous features. It quantifies the information certainty of each modal feature and adaptively allocates fusion weights, avoiding the problems of diluted reliable features and excessive amplification of noisy features in traditional fixed-weight fusion. The heterogeneous features of atmospheric waveguides (the electromagnetic characteristics of radar echoes and the physical characteristics of meteorological data) exhibit significant differences in reliability under different scenarios. For example, in near-sea evaporation waveguide regions, radar echoes can accurately capture abrupt changes in electromagnetic wave propagation (such as power attenuation jumps), but are susceptible to feature fluctuations due to sea surface clutter. Meteorological data (such as vertical temperature and humidity gradients), while directly reflecting the physical mechanism of waveguide formation, may suffer from missing local details due to sparse observation points. Feature entropy, as a quantitative indicator of information uncertainty, perfectly characterizes this reliability difference—the lower the entropy value, the more concentrated the feature distribution and the higher the certainty (such as stable meteorological gradient data); the higher the entropy value, the greater the feature fluctuation and the stronger the uncertainty (such as radar echoes interfered with by clutter).

[0069] Specifically, in step S4.2, the information entropy of the two modes in the interactive atmospheric waveguide map is calculated, the confidence level of the two modes is determined based on the information entropy of the two modes, and the aligned radar features and aligned meteorological features are adaptively fused according to the confidence level of the two modes to obtain the atmospheric waveguide prediction feature map. The adaptive fusion process is divided into three steps: First, calculate the radar cross-scale interactive feature sets separately. Interactive feature set with meteorological scale The information entropy is as follows:

[0070] in, The probability distribution of radar cross-scale interaction features. The information entropy of radar cross-scale interaction characteristics can be obtained similarly; for example, the information entropy of meteorological cross-scale interaction characteristics can be derived. Entropy directly reflects the intrinsic stability of each mode; Secondly, modal confidence is calculated based on entropy values:

[0071]

[0072] in, It is the confidence level of radar cross-scale interactive features. It is the confidence level of meteorological cross-scale interaction features; the entropy value is mapped to the weight of the 0-1 interval through the exponential function to ensure that low-entropy features (high certainty) receive high weight and high-entropy features (high uncertainty) receive low weight. Finally, dynamic fusion is achieved through the following formula:

[0073] in, and These are the aligned radar features and aligned meteorological features, respectively. This is an atmospheric waveguide prediction feature map.

[0074] This mechanism is highly adaptable to complex scenarios with non-uniform atmospheric waveguides: in areas of sudden change such as the passage of fronts, the temperature and humidity gradient entropy values ​​of meteorological data are low (the physical mechanism is clear). Automatic elevation prioritizes preserving global constraints of physical characteristics; in uniform areas such as open sea surfaces, the electromagnetic entropy value of radar echoes is low (propagation is stable). The corresponding enhancement strengthens the capture of local details; while in ambiguous areas where both modes have noise (such as the superposition area of ​​low signal-to-noise ratio radar echoes and sparse meteorological observations), the weights will automatically tend to be balanced, avoiding the dominance of single mode error in the fusion result. Through this entropy-driven dynamic weight allocation, the unique advantages of each mode can be preserved while suppressing mutual interference, and the final output is a fusion feature that combines physical consistency and detail accuracy, providing high-quality feature input for subsequent waveguide prediction.

[0075] In the construction of the loss function for non-uniform atmospheric waveguides, the total loss is adopted as a weighted sum of structural loss, classification loss and mode alignment loss. Its core design idea is to comprehensively improve the modeling accuracy of complex waveguide features through the collaborative optimization of multi-dimensional constraints—ensuring the physical consistency of spatial structure, strengthening the ability to identify waveguide types, and eliminating the systematic bias of heterogeneous mode fusion.

[0076] First, structural loss ( This study focuses on the spatial morphology reconstruction of non-uniform waveguides. By calculating and predicting the similarity between waveguides and real waveguides in structural features such as gradient distribution, boundary contours, and regional continuity, the model is constrained to accurately capture microscale abrupt changes (such as thickness jumps in evaporating waveguides) and macroscale trends (such as the extension direction of frontal waveguides). For regions with significant non-uniformity (such as waveguide gradient bands at the sea-land interface), structural loss can effectively avoid "fuzzy boundaries" or "false continuity" in the prediction results, ensuring that the spatial distribution of waveguides conforms to the physical laws of electromagnetic propagation.

[0077] Secondly, classification loss ( This approach is designed for fine-grained differentiation of waveguide types. It uses cross-entropy to quantify the prediction errors of different types of waveguides, such as evaporating waveguides, surface waveguides, and suspended waveguides. In non-uniform environments, the characteristics of different types of waveguides often overlap (e.g., the refractive index profiles of low-height surface waveguides and evaporating waveguides are similar). The classification loss enhances inter-class differences and intra-class compactness, driving the model to learn type-specific discriminative features (e.g., the "top gradient abrupt change" feature of suspended waveguides), thereby reducing type misclassification.

[0078] Finally, the modal alignment loss ( As a key constraint for heterogeneous feature fusion, the maximum mean difference (MMD) is used to measure the distribution deviation between radar electromagnetic features and meteorological physical features in high-dimensional space, ensuring that the two modes have consistent statistical properties during fusion. In non-uniform scenarios, radar echoes may generate noise due to multipath effects, and meteorological data may have sampling biases due to sparse observations. The mode alignment loss is corrected by backpropagation to correct the feature mapping parameters, avoiding "fusion offset" caused by mode distribution imbalance (such as over-reliance on noisy radar features while ignoring key meteorological gradients).

[0079] The three losses are weighted by coefficients ( , , Dynamic equilibrium: In regions with complex waveguide structures but a single type (such as the gradient transition region of a large-scale surface waveguide), increasing... Prioritize ensuring structural restoration; in non-uniform zones with mixed waveguide types (such as frontal regions where multiple waveguides intersect), improve... Enhance type differentiation; in scenarios with significant modal noise (such as areas where low signal-to-noise ratio radar data and sparse meteorological observations overlap), add... This enhances alignment constraints. This weighted collaborative mechanism enables the total loss to adapt to diverse scenarios of non-uniform atmospheric waveguides, ultimately driving the model output to provide predictions that combine structural realism, type accuracy, and modal consistency.

[0080] Therefore, the total loss is a weighted sum of the structural loss, classification loss, and modal alignment loss:

[0081] in, , , These are the weighting coefficients; For structural similarity loss; This represents the cross-entropy loss.

[0082] To reduce model parameter adjustments during training, the loss structure loss and modality alignment loss for regression problems are defined as regression loss, and the classification loss is defined as classification loss. The optimized loss is then weighted and optimized as follows:

[0083] in, For regression loss weighting factors, This is the regression loss weighting factor.

[0084] like Figure 2 As shown, to verify the prediction accuracy, this embodiment is based on simulation data and generated using an atmospheric refractive index theoretical model. The specific parameters are as follows: Waveguide type and thickness: It covers three typical atmospheric waveguides, namely evaporation waveguides (thickness 0-50m, accounting for 40%), surface waveguides (thickness 50-500m, accounting for 35%), and suspended waveguides (thickness >500m, accounting for 25%). In a single sample, the coexistence of multiple waveguide types accounts for 30% of the scenarios (reflecting non-uniformity).

[0085] Refractive index gradient (unit: N units / 100m, 1N=10) -6 The value range is -150 to 50 N units / 100m, and the gradient in the waveguide region is generally <-40 N units / 100m (negative gradient). Sample quantity and distribution: A total of 1200 valid samples were generated, including 450 non-uniform mutation samples (waveguide thickness / gradient mutation amplitude >20%), 320 similar morphology samples (similarity of refractive index profiles of different waveguide types >80%), and the rest were regular uniform samples. The samples were divided into a training set (840 sets) and a test set (360 sets) in a 7:3 ratio.

[0086] Figure 2 The key performance indicators for atmospheric waveguide prediction using different methods are compared, and the coordinate axes and curve meanings are as follows: Horizontal axis (X-axis): Atmospheric waveguide refractive index gradient (unit: N, 1N=10⁻⁶ units) -6 The value ranges from 16 to 34N, covering three non-uniform scenarios: weak gradient (16-22N), medium gradient (22-28N), and strong gradient (28-34N). The larger the gradient, the more complex the waveguide structure and the stronger the abrupt change. Vertical axis (Y-axis): Comprehensive evaluation index (after normalization), with a value range of 15-55, obtained by weighted calculation (the weights are set based on the actual needs of waveguide prediction). The larger the value, the better the comprehensive prediction performance. The meaning of the four curves: Raw data: Measured index curves of real atmospheric waveguides, representing the "ideal prediction target" and serving as the benchmark for comparison among all methods; Fine-grained classification algorithm: based on the static graph convolution method for fine-grained vehicle classification (the second type of prior art mentioned in the background); SP-URNet: A prediction method based on 3D convolutional autoencoder and TCN (the mainstream existing technologies mentioned in the background). This invention: A method based on non-uniform feature fusion and graph inference (the technical solution proposed in this paper).

[0087] Based on experiments, Figure 2 The experimental results and data comparison show that the present invention, through the technology of dynamic topological feature encoding, hierarchical graph construction, cross-scale graph inference and modality alignment aggregation, significantly outperforms existing methods in three core dimensions: small-scale feature preservation, cross-scale association capture, and similar type differentiation. It also has the highest degree of fit with the original data, especially in strong gradient non-uniform scenarios. This provides higher-precision technical support for atmospheric waveguide environment prediction for maritime over-the-horizon communication and target detection.

[0088] Example 2 This embodiment provides an atmospheric waveguide prediction system based on non-uniform feature fusion and graph derivation, including: The dynamic topology feature encoding module is used to perform dynamic topology feature encoding based on the radar echo gradient change rate and the intensity of meteorological feature disturbances to obtain multimodal reconstruction features; The layered graph construction module is used to construct corresponding graph structures based on different scales of multimodal reconstructed features, and to calculate the edge weights of graph structures at each scale by integrating spatial distance and feature similarity to form an atmospheric waveguide graph. The cross-scale graph inference module is used to determine whether to interact with nodes of the target scale and nodes of other scales in the atmospheric waveguide map. For all nodes of the target scale and nodes of other scales that are determined to interact, cross-scale interaction is performed based on a gating mechanism to generate an interactive atmospheric waveguide map. The feature calibration and aggregation module is used to perform mode alignment on the cross-scale interactive features of two modes in the interactive atmospheric waveguide map, and to perform adaptive aggregation based on the confidence between the two modes of cross-scale interactive features to obtain the atmospheric waveguide prediction feature map.

[0089] The examples and application scenarios implemented by the above modules and corresponding steps are the same, but are not limited to the content disclosed in Embodiment 1 above. It should be noted that the above modules, as part of the system, can be executed in a computer system such as a set of computer-executable instructions.

[0090] The descriptions of each embodiment in the above embodiments have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0091] The proposed system can be implemented in other ways. For example, the system embodiments described above are merely illustrative, and the division of modules described above is only a logical functional division. In actual implementation, there may be other division methods. For example, multiple modules may be combined or integrated into another system, or some features may be ignored or not executed.

[0092] Example 3 This embodiment provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps in the atmospheric waveguide prediction method based on non-uniform feature fusion and graph deduction as described in Embodiment 1 above.

[0093] Example 4 This embodiment provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the steps in the atmospheric waveguide prediction method based on non-uniform feature fusion and graph derivation as described in Embodiment 1 above.

[0094] Example 5 This embodiment provides a computer program product or computer program, including computer instructions stored in a computer-readable storage medium. The processor of a computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computer device to perform the steps in the atmospheric waveguide prediction method based on non-uniform feature fusion and graph derivation described in Embodiment 1 above.

[0095] Those skilled in the art will understand that embodiments of the present invention can provide methods, systems, or computer program products. Therefore, the present invention can take the form of hardware embodiments, software embodiments, or embodiments combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage and optical storage) containing computer-usable program code.

[0096] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, as well as combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0097] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0098] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0099] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. The storage medium can be a magnetic disk, optical disk, read-only memory (ROM), or random access memory (RAM), etc.

[0100] While the specific embodiments of the present invention have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. A method for atmospheric duct prediction based on non-uniform feature fusion and graph inference, characterized in that, The method comprises the following steps: According to the radar echo gradient change rate and the meteorological feature disturbance intensity, dynamic topological feature coding is performed to obtain multi-modal reconstruction features; According to the different scales of the multi-modal reconstruction features, corresponding graph structures are constructed, the edge weights of the graph structures of each scale are calculated by fusing the spatial distance and the feature similarity, and an atmospheric waveguide graph is formed; In the atmospheric waveguide graph, whether to interact is determined according to the edge weights of the nodes of the to-be-interacted scale and the nodes of other scales, and for all the nodes of the to-be-interacted scale and the nodes of other scales that are determined to interact, cross-scale interaction is performed based on a gating mechanism to generate an interactive atmospheric waveguide graph; The two kinds of modal cross-scale interaction features in the interactive atmospheric waveguide graph are aligned, and the atmospheric waveguide prediction feature map is obtained by adaptively aggregating according to the confidence between the two kinds of modal cross-scale interaction features.

2. The atmospheric duct prediction method based on non-uniform feature fusion and graph deduction according to claim 1, characterized in that, According to the radar echo gradient change rate and the meteorological feature disturbance intensity, dynamic topological feature coding is performed to obtain multi-modal reconstruction features, including: Based on the radar echo graph and the meteorological observation dataset, the radar echo gradient change rate and the meteorological feature disturbance intensity are calculated; When the radar echo gradient change rate exceeds the electromagnetic characteristic threshold or the meteorological feature disturbance intensity exceeds the physical characteristic threshold, it is determined to start topological reconstruction; The electromagnetic feature tensor of the radar echo graph and the physical feature tensor of the meteorological observation dataset are extracted, and a topological weight matrix is generated after processing by a spatio-temporal attention mechanism; According to the topological weight matrix and the adjusted node connection relationship, the multi-modal reconstruction features are generated by a parameterized convolution kernel generation function.

3. The atmospheric duct prediction method based on non-uniform feature fusion and graph deduction according to claim 1, characterized in that, According to the different scales of the multi-modal reconstruction features, corresponding graph structures are constructed, the edge weights of the graph structures of each scale are calculated by fusing the spatial distance and the feature similarity, and an atmospheric waveguide graph is formed, including: According to the different scales of the multi-modal reconstruction features, the features are divided into microscale, mesoscale and macroscale; The spatial distance and the feature similarity of the multi-modal reconstruction features in different scales are fused to calculate the edge weights of the corresponding scale graph structure; Based on the edge weights of the graph structures of different scales, each multi-modal reconstruction feature is taken as a node, and the correlation between the nodes is taken as an edge to construct the graph structures of different scales, and the atmospheric waveguide graph is formed by using the graph structures of different scales.

4. The atmospheric duct prediction method based on non-uniform feature fusion and graph deduction according to claim 1, characterized in that, In the atmospheric waveguide graph, whether to interact is determined according to the edge weights of the nodes of the to-be-interacted scale and the nodes of other scales, and for all the nodes of the to-be-interacted scale and the nodes of other scales that are determined to interact, cross-scale interaction is performed based on a gating mechanism to generate an interactive atmospheric waveguide graph, including: In the atmospheric waveguide graph, if the edge weights of the nodes of the to-be-interacted scale and the nodes of other scales exceed the weight threshold, the two nodes interact; For all the nodes of the to-be-interacted scale and the nodes of other scales that are determined to interact, the corresponding cross-scale interaction features are calculated based on the gating mechanism, and the interactive atmospheric waveguide graph is generated by combination.

5. The atmospheric duct prediction method based on non-uniform feature fusion and graph deduction according to claim 1, characterized in that, The two kinds of modal cross-scale interaction features in the interactive atmospheric waveguide graph are aligned, and the atmospheric waveguide prediction feature map is obtained by adaptively aggregating according to the confidence between the two kinds of modal cross-scale interaction features. The two kinds of cross-scale interaction modal features in the interaction atmospheric duct graph are mapped to the same high-dimensional space, and when the maximum mean difference of the two kinds of cross-scale interaction modal features in the high-dimensional space meets a difference threshold, the aligned radar features and the aligned meteorological features are obtained. The information entropy of the two kinds of cross-scale interaction modal features in the interaction atmospheric duct graph is calculated, the confidence of the two kinds of cross-scale interaction modal features is determined based on the information entropy of the two kinds of cross-scale interaction modal features, the aligned radar features and the aligned meteorological features are adaptively fused according to the confidence of the two kinds of cross-scale interaction modal features, and the atmospheric duct prediction feature graph is obtained.

6. The atmospheric duct prediction method based on non-uniform feature fusion and graph deduction according to claim 3, characterized in that, The edge weight of the corresponding scale graph structure is calculated as follows: where the edge weight , the spatial decay term , is the scale lower vertex and spatial distance, is the cosine similarity, is the decay coefficient, is the summation sign.

7. The atmospheric duct prediction system based on non-uniform feature fusion and graph inference, characterized in that, The method comprises the steps of: The dynamic topological feature coding module is configured to perform dynamic topological feature coding according to the radar echo gradient change rate and the meteorological feature disturbance intensity, and obtain multi-modal reconstruction features. The hierarchical graph construction module is configured to construct corresponding graph structures according to different scales of the multi-modal reconstruction features, fuse spatial distance and feature similarity to calculate edge weights of the graph structures of each scale, and form an atmospheric duct graph. The cross-scale graph reasoning module is configured to determine whether to interact according to the edge weights of the nodes of the to-be-interacted scale and the nodes of other scales in the atmospheric duct graph, and perform cross-scale interaction based on a gating mechanism for all to-be-interacted nodes of the to-be-interacted scale and the nodes of other scales that are determined to interact, to generate an interaction atmospheric duct graph. The feature calibration and aggregation module is configured to perform modal alignment on the two kinds of cross-scale interaction modal features in the interaction atmospheric duct graph, and perform adaptive aggregation according to the confidence between the two kinds of cross-scale interaction modal features, to obtain an atmospheric duct prediction feature graph.

8. A computer-readable storage medium having stored thereon a computer program, characterized in that, The program is executed by the processor to implement the steps in the atmospheric duct prediction method based on non-uniform feature fusion and graph reasoning in any one of claims 1-6.

9. A computer device comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, The processor executes the program to implement the steps in the atmospheric duct prediction method based on non-uniform feature fusion and graph reasoning in any one of claims 1-6.

10. A computer program product, characterised in that, The computer program product comprises a computer program which, when executed by a processor, implements the steps in the atmospheric duct prediction method based on non-uniform feature fusion and graph reasoning in any one of claims 1-6.

Images